https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125476
--- Comment #5 from Robin Dapp <rdapp at gcc dot gnu.org> --- (In reply to Zhongyao Chen from comment #4) > For this corner case, I think either result is fine. Scalar code should > not be worse here, since scalar and vector throughput are about the > same. > > These tests are sensitive to cost model changes though. How should we > handle regression like this? I also tried -fno-vect-cost-model, but that did > not help here. Does adjusting the params help (--param=gpr2vr-cost=1 --param=vr2gpr-cost=1)? We added them specifically for these test cases.
